AI assistant integration

Run Notra from your AI assistant.

The Notra MCP server lets Claude Desktop — or any MCP client — start audits, poll them, and read verified findings directly in a conversation. Ask it “what's exposed on staging?” and it runs the audit and answers with evidence, not vibes. Same API key, same limits, no special access.

Six tools, same contract

Ask, don't click

Ask your assistant “what's exposed on staging?” and it can start the audit, wait for the verdict, and read back verified findings — each with the exact request/response that proved it — without you leaving the conversation.

Gate in conversation

The check_gate tool returns the same pass/fail verdict the CI gate uses, so “would staging pass a fail-on-high gate right now?” gets a real answer instead of a guess.

Same leash as every key

The MCP server has no special access of its own — same bearer key, same org scoping, same rate limits, same ownership checks as every other API request.

The full tool list — start_scan, get_scan_status, get_findings, check_gate, get_report_url, list_scans — is documented in the MCP docs, and each tool maps to the REST endpoints in the API reference.

What pointing your assistant at it looks like

One small server file, one config entry. The setup — download, dependencies, and the client config for Claude Desktop — is a two-minute job documented step by step in the MCP docs; the shape is:

claude_desktop_config.json — illustrativefull setup in the docs

{
  "mcpServers": {
    "notra": {
      "command": "node",
      "args": ["/absolute/path/to/notra-mcp/notra-mcp.mjs"],
      "env": {
        "NOTRA_API_KEY": "your dashboard API key",
        "NOTRA_API_BASE": "https://your-notra-domain"
      }
    }
  }
}

Scope, honestly

The integrations that exist today are exactly three: this MCP server, the REST API behind it, and the CI/CD gate built on the API (see GitHub Actions and GitLab CI). There is no chat-tool or ticketing bridge yet — findings arrive in the conversation, the dashboard report, and the API, and if you want them piped somewhere else, the API is the contract to build that on. The server file itself is small enough to read before you trust it with a key.

Ownership still applies

The assistant can only scan domains your org has proven it owns. Asking it to audit something else gets the same refusal the dashboard would give you.

Evidence travels with findings

Every finding the assistant reads back carries the exact request/response that proved it — same evidence standard as the report.

Revoke is one act

Delete or rotate the API key and the assistant's access ends with it. There is no separate AI credential to track down.

Questions before you connect it

Does the assistant get more access than an API key?

No. The MCP server is a thin adapter over the same REST API endpoints every other integration uses — same bearer key, same org scoping, same rate limits, same ownership checks. The server has no special access of its own, so revoking the key revokes the assistant.

Which clients work with it?

Any MCP client — Claude Desktop is the primary one, but any client that speaks MCP can load the server and get the same six tools. Setup is a small Node script and a client config entry, both documented in the MCP docs.

Can it start a scan on any domain?

No. The domain must already be ownership-verified for your org, the same gate as a dashboard scan or an API call — an unverified target is refused. Your assistant can only audit what your org has proven it owns.

Not sure it's worth wiring up?

Run the free scorecard first. It costs nothing, takes about a minute, and tells you whether a verified audit is worth your attention.

Get your free scorecard